One model, served seriously.
Versioned. Metered. OpenAI-compatible.
WebikAI Inference serves exactly one open-weight model — KAT-Coder V2.5 — behind the standard OpenAI API. Every response names the immutable release that produced it, and billing is a prepaid wallet metered from engine-reported tokens. Nothing changes under you unannounced.
curl https://api.webik.ai/v1/chat/completions \
-H "Authorization: Bearer wk_live_..." \
-H "Content-Type: application/json" \
-d '{
"model": "kwaipilot/kat-coder-v2.5-dev",
"messages": [{"role": "user", "content": "Refactor this function"}]
}'- model
- kwaipilot/kat-coder-v2.5-dev
- system_fingerprint
- wbk_8d3d0b9a7f2c
- x-webik-model-release
- webik/kat-coder-v2.5-dev-2026-08-w4a16-r1
- usage
- 1280 in · 412 out
- request_id
- req_01j...
Every response names the exact serving release.
Request the canonical ID kwaipilot/kat-coder-v2.5-dev, or pin an
immutable release ID (the form is webik/kat-coder-v2.5-dev-2026-08-w4a16-r1) when a build must never move under you.
No silent model changes.
Every build of the model is an immutable, dated release, and every response names the one that served it. A new upstream version is a new canonical ID — never a quiet swap behind the name you pinned.
How releases work →Prepaid, fail-closed billing.
You load a wallet; requests settle from engine-reported usage. The balance never goes negative, capacity errors never debit, and there is no subscription to forget about.
See pricing →A gated serving contract.
Published context, output, and capability limits — and a deprecation policy with dated notices and sunset headers. Prices finalize only when the launch performance gate passes.
Read the docs →Serious inference for one serious model.
The founding beta is 25 invited accounts running real coding-agent workloads. Prepaid from $20, introductory per-token pricing, and a direct line to the people running the fleet.
Apply for the founding beta